Weighted Dynamic Brain Network Feature Analysis in Schizophrenia

Abstract

Most existing schizophrenia studies employ graph neural network classification based on static brain network structures, neglecting temporal dynamics. Alternatively, they typically introduce dynamic elements solely at the node feature or graph structure level, thereby limiting the models' expressive capabilities. To fully explore the spatio-temporal characteristics of EEG signals, this paper proposes a temporal dynamic graph convolutional network (TDGCN) based on temporal features. It simultaneously introduces temporal dynamic modeling mechanisms at both the node feature and graph structure levels. Wavelet packet-based statistical features and LSTM are used by the model to generate time-varying dynamic weight. Dynamically optimized graph structures and node features are fed into the GCN by the model, and then features are adjusted through a channel gating mechanism after the convolutional layer. Thus, the approach facilitates efficient modeling and classification of the dynamic evolution of brain regions. Experiments on the first-episode schizophrenia resting-state EEG dataset demonstrate that TDGCN achieves optimal performance in the gamma band, achieving an accuracy of 93.85% and outperforming baseline models across multiple evaluation metrics. Ablation experiments validated the crucial role of temporal dynamic weight and dual-layer dynamic graph modeling, demonstrating TDGCN's ability to effectively capture the spatiotemporal non-stationarity of brain networks in schizophrenia patients. This offers novel insights for intelligent auxiliary diagnosis of psychiatric disorders.